On Tue, 22 Apr 2003 14:03:33 -0400
Ravi Varadhan <rvaradha@jhsph.edu> wrote:
> Dear Peter:
>
> There is nothing special to program in propensity score models. You simply
> use existing logistic regression ("glm" with logit or some other links)
> software for a binary treatment variable, as a function of all covariates
> that may possibly affect the assignement. In other words, you model
> Pr(Z=1 | covariates), where Z=1 indicates treatment and Z=0 indicates
> placebo or some other treatment.
>
> Ravi.
In addition to what Ravi said, in invaluable tool is lowess with the iter=0
option (or use the plsmo function in the Hmisc library which uses this) if your
outcome variable is binary. Stratify by actual treatment received and get a
nonparametric estimate relating propensity for treatment to probability of
outcome using lowess. This does not adjust for subject heterogenity (hence
odds ratios are biased towards the null) as you do in the final model (which
includes propensity and covariates) but it adjusts for confounding on a
continuous basis. In my experience this is better than adjusting just for
quintiles of propensity. If your outcome is continuous and ols is appropriate,
you can use lowess or loess with default parameters to get this graph. These
graphs do not assume a functional form for propensity vs. outcome, only
smoothness.
Continuous analyses are greatly preferred but to demonstrate the adjustment for
confounding to non-statisticians, matched-sets analyses are educational,
matching on propensity. Some functions in Hmisc help with this.
---
Frank E Harrell Jr Prof. of Biostatistics & Statistics
Div. of Biostatistics & Epidem. Dept. of Health Evaluation Sciences
U. Virginia School of Medicine http://hesweb1.med.virginia.edu/biostat
|